Convert any PDF or PPTX deck into a self-contained interactive HTML deck, one file, one link, shareable. Auto-detects input file extension (.pdf or .pptx) and routes to the matching extraction path. Detects which of the 5 deck types the source is (pitch, sales, launch, keynote, all-hands) and hands off to the matching deck-* skill. Trigger when the user uploads a .pdf or .pptx of any deck type, or says any variation of "convert this PDF to HTML", "turn this PPTX deck into HTML", "import this deck", "upgrade my deck", "make my deck interactive", "migrate my deck from PDF/PPTX". Mirrors source palette, typography, and slide structure 1-to-1 (no redesign), preserves slide count, uses extracted screenshots on demo slides where applicable, auto-extracts key fields, then confirms with the user before building. Reuses deck-builder core pipeline and per-type spines as role-labeling taxonomy. Fastest path from cold PDF or PPTX to warm shareable link.
Convert any PDF or PPTX deck into a self-contained interactive HTML deck, one file, one link, shareable. Auto-detects input file extension (.pdf or .pptx) and routes to the matching extraction path. Detects which of the 5 deck types the source is (pitch, sales, launch, keynote, all-hands) and hands off to the matching deck-* skill. Trigger when the user uploads a .pdf or .pptx of any deck type, or says any variation of "convert this PDF to HTML", "turn this PPTX deck into HTML", "import this deck", "upgrade my deck", "make my deck interactive", "migrate my deck from PDF/PPTX". Mirrors source palette, typography, and slide structure 1-to-1 (no redesign), preserves slide count, uses extracted screenshots on demo slides where applicable, auto-extracts key fields, then confirms with the user before building. Reuses deck-builder core pipeline and per-type spines as role-labeling taxonomy. Fastest path from cold PDF or PPTX to warm shareable link.
Deck Import (PDF or PPTX, HTML Deck, generic across 5 deck types)
Keep your process invisible. Do not narrate the skill architecture (the "type pack" and "core" split) or announce which files you are reading. Skip preambles like "I'll start by reading...". Gather what you need, then build the deck.
You are converting a PDF or PPTX deck into an interactive HTML deck, the entry point for users who already have a deck and want the upgrade (shareable link, scale-to-fit canvas, navigation shell) without a redesign.
Core principle: the user's deck is already the deck. Your job is to faithfully reproduce it in HTML, not to redesign it. Mirror the palette. Mirror the typography where possible. Preserve slide count and order. Keep their words. The "upgrade" is the delivery medium, not the content.
Two build modes, page-image is the default. The skill ships two Phase 3 implementations:
Mode A, page-image (DEFAULT): each slide is rendered as a full-bleed background image of the source page. The HTML provides only the shell, role-meta pill on interior slides, progress bar, nav arrows, counter, scale-to-fit canvas, touch swipe. Pixel-faithful by construction, no palette drift, no typography mismatch, no source-text ghosting, no logo recreation. One-shot quality on virtually every deck. See references/page-image-mode.md.
Mode B, reconstruction (OPT-IN): every slide is rebuilt in HTML/CSS/SVG. Editable, but risk-prone: palette guessing, source-text ghosting on photo backgrounds, content rewriting, layout drift. Use only when the user explicitly asks to be able to edit slides later, or to upgrade specific visual components. The user opts in during Phase 2 by replying "reconstruction" or "rebuild".
Default Mode A. Switch only on explicit opt-in.
This skill is generic across 5 deck types, pitch, sales, launch, keynote, all-hands. It is NOT a general document converter. PDFs or PPTX files that aren't decks (whitepapers, resumes, research papers, portfolios) should be redirected to a doc-builder skill.
Conceptually, deck-import is a type-pack-like layer in front of deck-builder + the matching deck-<type> skill (Mode B only, Mode A doesn't need the handoff skill for build, only for role-label vocabulary). Mode A adds three front-end phases (Extract, Analyze, Confirm) plus a thin Mode A shell build. Mode B adds the same front-end phases then hands off to the standard reviewer gates per references/review-adaptations.md.
If the user has not installed deck-builder or the relevant deck-<type> skill for the detected type, surface the gap early and ask them to install.
Auto-detect (Phase, 1, file-extension routing)
Before any extraction runs, examine the input file extension:
input_path = <user-supplied file path>
ext = lowercase suffix of input_path
if ext == ".pptx":
→ PPTX-only path (Phase 0p)
elif ext == ".pdf":
→ PDF path with optional sibling .pptx check (Phase 0)
else:
→ ask the user: "deck-import handles .pdf and .pptx files only. What did you mean?"
The PPTX-only path skips PDF rasterization entirely and uses scripts/extract-pptx-only.py (see Sub-phase C). It produces a JSON manifest with per-slide text, embedded images, speaker notes, and slide order. The PDF path runs the established extraction pipeline and OPTIONALLY uses sibling PPTX assets when available for higher fidelity.
Both paths converge at Phase 1 (Analyze) using the same classification.json shape, so everything downstream (Confirm, Build, Review, Release, Learn) is identical regardless of input format.
When this skill runs (trigger anatomy)
The trigger description (in YAML frontmatter above) is deliberately broad because this is the primary adoption ramp for users who already have a deck. If the user uploads a PDF or PPTX that could plausibly be any of the 5 deck types AND mentions HTML, interactive, conversion, import, migration, or "upgrade" in any form, this skill is right.
Edge cases:
File ambiguous between two deck types (e.g., reads like an all-hands OR a keynote): the Phase 1 classifier surfaces both in the confirmation block with scores. The user picks via "this is a " in Phase 2.
File is a research paper, whitepaper, or article: this is NOT the right skill. Redirect to a doc-builder.
File is a resume or portfolio: not the right skill. Redirect to the appropriate doc skill.
File is a handbook, wiki export, or multi-page report: not a deck. Redirect to a doc-builder.
The classifier treats "is this a deck at all?" as a sanity check. If no type scores above the "low" confidence threshold AND the page count is outside every canonical band (e.g., 40+ pages with dense prose), surface that observation and ask the user whether they want to proceed anyway, re-route to a doc-builder, or cancel.
The conversion pipeline (6 phases)
Phase
Owner
Gate
Output
0. Extract
script (extract_deck_pdf.py or extract-pptx-only.py)
script exits 0
JSON dump of per-slide text, per-slide images, metadata
The deck-builder 5-phase core maps onto Build, Review, Release, Learn (phases 3 to 6 here). Phases 0 to 2 (Extract, Analyze, Confirm) are this skill's original contribution.
Phase 0 , Extract (PDF path)
Input: the PDF path (from upload or user-supplied filepath).
Output: a JSON file at /tmp/deck-import-<timestamp>/extraction.json plus a per-page images folder at /tmp/deck-import-<timestamp>/pages/ (one PNG per slide). When PPTX is available (see Step 0a), an additional pptx_assets.json and an unzipped ppt/ tree.
Gate: extraction script exits cleanly. If it doesn't, ask the user what's different about this PDF (encrypted? scanned? very large?).
Step 0a , PPTX-first check (prefer PPTX as the primary source when available)
Before running the PDF extractor, check whether a .pptx of the same deck is reachable. PPTX gives roughly 5x the source fidelity of PDF-only extraction (exact source PNGs in ppt/media/, exact gradient color stops from a:gradFill, exact EMU positions from p:spPr/a:xfrm, exact a:custGeom shapes for non-standard outlines like semicircles or half-domes). Conversion runs that started PDF-only and only surfaced the PPTX mid-run consistently spend 6 to 8 extra iteration rounds reverse-engineering rasters and geometries that the PPTX already contains exactly.
Decision tree:
Look for a sibling .pptx in the same directory as the uploaded .pdf (same stem; e.g., MyDeck.pdf next to MyDeck.pptx).
If not present, look in the user's workspace folder for a .pptx with a matching stem.
If still not found, ask the user before iterating from PDF: "Do you have a .pptx of this deck? PPTX gives much more accurate extraction (exact source images, gradients, positions) and is the difference between 4 rounds and 12 rounds on most decks. If yes, drop it in and tell me the name."
If a PPTX is found, run BOTH extractors and treat the PPTX as the primary source for assets, the PDF as the rasterization fallback:
extract_pptx_assets.py enumerates ppt/slides/*.xml plus their relationship files and emits a JSON document with, per slide: a list of pic refs (rId, image_path, position_px, size_px) keyed back to the unzipped media folder, plus a list of a:gradFill records (color_stops, angle). Phase 1 reads this JSON for geometry and gradient ground truth.
Step 0b , Read slide size first; cache (sldW_emu, sldH_emu) before any EMU-to-pixel conversion
Whether or not PPTX is available, but mandatory when PPTX is the source: parse ppt/presentation.xml and cache the <p:sldSz cx="..." cy="..."/> values before doing any EMU-to-pixel math. Default PowerPoint slide sizes range from 10x7.5" (4:3, cy=6858000) through 10x5.625" (16:9, cy=5143500) through 13.333x7.5" (16:9 widescreen, cy=6858000 cx=12192000). Assuming the default 4:3 dimensions when the deck is 16:9 puts every Y position off by roughly 33%, components land below the intended row, headers drift up, bottoms get clipped.
The cached (sldW_emu, sldH_emu) becomes the divisor in every later position/size computation: position_px = emu / sldW_emu * 1440 for x, emu / sldH_emu * 810 for y. Bake this into the build script as constants at the top, not as inline literals.
Step 0c , PDF extraction (always run; primary when PPTX unavailable, fallback rasterization when PPTX is the source)
pages/page-<N>.png per slide (at 2x resolution for high-quality screenshot placement)
If the PDF is scanned (no text layer), the script falls back to OCR via pytesseract. See references/extraction-pipeline.md for the full data contract and fallback rules.
Before Phase 1 finishes, scan the user's workspace for prior imports of the same company or the same shell signature. These are the canonical style reference, match them, don't reinvent.
ls <user-workspace>/*-Imported.html 2>/dev/null
ls <user-workspace>/*-Imported.html 2>/dev/null | xargs -I{} grep -l "data-shot" {} | head -5
Two signals to surface in the Phase 2 confirmation block:
Company match. Any prior *-Imported.html whose filename includes the detected company (case-insensitive). Display: Prior imports, <files>. Match their style? (yes/no, default yes).
Shell-signature match. Any prior import using a <div class="slide-img" data-shot="N"> pattern indicates Mode A page-image. Surface as: Prior decks use Mode A page-image, defaulting to same.
Phase 0p , Extract (PPTX-only path)
Input: the PPTX path (from upload or user-supplied filepath).
Output: a JSON file at /tmp/deck-import-<timestamp>/pptx-manifest.json plus an extracted images folder at /tmp/deck-import-<timestamp>/media/ (one image per embedded asset).
Gate: extraction script exits cleanly.
media/image-<slide>-<idx>.<ext> per embedded image, with predictable naming
Speaker notes are preserved verbatim from each slide's notes pane. Slide order is the raw presentation.xml order, not filename-sorted (PowerPoint reorders slides without renumbering filenames).
The manifest shape was designed so Phase 1 (classify_slides.py) can consume it directly without translation, by treating each slide as if its body_text were the PDF page's extracted text. Image-position math uses the same EMU divisors as the PDF+PPTX hybrid path (see Step 0b).
PPTX-only skips Step 0a (no PDF to consult), Step 0c (no PDF extraction), and the page-image rasterization for Mode A. To produce Mode A page-images from a PPTX-only input, the build step converts each slide to PNG via LibreOffice headless conversion (libreoffice --headless --convert-to png --outdir <out> <pptx>) when LibreOffice is installed; otherwise the build step asks the user to provide a PDF render or falls back to Mode B reconstruction.
Phase 1 , Analyze
Input: extraction.json (PDF path) or pptx-manifest.json (PPTX-only path), plus per-slide images.
Output: classification.json + a draft conversion brief (in memory / shown to user in Phase 2).
Gate: none, this phase is non-destructive analysis.
Analysis runs in two steps, both driven by scripts/classify_slides.py:
Step 1: Detect deck type
The classifier scores all 5 deck types by:
Deck-wide match against each type's type-signature keywords (weighted phrases relatively unique to that type, e.g., "use of funds" signals pitch, "available now" signals launch, "Q&A" signals all-hands).
Slide-count fit bonus: how close page_count is to each type's canonical band (pitch is roughly 14, sales is roughly 11, keynote is roughly 28, launch is roughly 12, all-hands is roughly 15).
Confidence:
high: winner >= 15 score AND >= 1.8x runner-up.
medium: winner >= 8 AND >= 1.3x runner-up.
low: otherwise. Classifier emits the top-3 alternatives so the confirmation block can show them and let the user override.
The winning type's matching deck-<type> skill is the handoff skill for Phase 3. Full vocabulary: references/spine-tables.md.
Icon classification (when PPTX present)
For each icon SVG/PNG in ppt/media/, classify intended pairing:
icon-on-colored-square, used inside .stat-icon, .sol-icon, .wn-icon, .bm-feat .ic, .vis-st-row .ic, etc. Source fill is #ffffff or #fdfdfd. Render as WHITE.
icon-on-light-bg, standalone outline icon on a light slide background. Source fill is #006838 or #000000 (or the deck's primary). Render as PRIMARY color.
Map per-slot in the conversion brief. Before ImageMagick conversion:
The inherit fill is unsupported by ImageMagick and must be stripped, leaving it in produces transparent glyphs even when the fill cascade should resolve. Detailed recipe in references/extraction-pipeline.md "PPTX icon color normalization".
Logo stack detection
When the source has a logo wall with stacked logos, detect pairing by overlapping x-ranges in bounding boxes, NOT by left-to-right adjacency. Phase 1 emits (x, y, w, h) per logo; group into vertical stacks where two logos' x-ranges overlap and their y-ranges differ by more than half the canvas-height of the wall region.
Horizontal-adjacency grouping pairs logos by reading order, which is wrong when the source visually stacks them. The fix matches what the user's eye sees. Recipe in references/extraction-pipeline.md "Logo stack detection".
Step 2: Classify slide roles (using the winning type's spine)
For each slide, detect:
Role (a spot on the winning type's spine): keyword-match the extracted text against the spine vocabulary in references/spine-tables.md. Confidence levels: high / medium / low. Low-confidence slides are labeled "Content" and flagged for user correction in the confirmation block. Hard overrides still apply (Cover = page 1; Team-like slide = 3+ human names + role title; Ask/Asks = dollar amount + raise language, each adapted to which labels the winning spine actually uses).
Company / counterparty name (from title slide text + metadata title).
Type-dependent context (skill populates only fields matching the detected type):
pitch, stage + sector
sales, industry + deal-size band
keynote, venue + talk length
launch, product + availability
all-hands, size + meeting cadence
Dominant palette (3 to 5 colors): sample each page image with detect_palette.py. Cross-page aggregation to identify the true brand primary vs. per-slide accents. See references/palette-typography-detection.md.
Typography family guess: inspect pdfplumber font metadata per text block (PDF path) or python-pptx font metadata (PPTX-only path). Most files expose the actual font names (Inter-Regular, Helvetica, TiemposHeadline-Medium). When a detected font isn't web-safe, map to the closest Google Font stand-in per the table in references/palette-typography-detection.md.
Images per slide: note which slides have full-bleed or hero imagery. These become static screenshots in the HTML. For types with a Demo spine role (pitch, launch), extract the largest image on the detected demo slide as the hero and label it "[Static product preview, interactive demo available on request]". For all other types, imagery is handled per the handoff skill's visual patterns.
Missing-spine-role flags: if the winning type's canonical spine has roles without a matching slide, note for the confirmation block. This is advisory, not prescriptive, we preserve the source structure 1:1.
Image content classification: tag every extracted image as photographic-real / ai-stylized / illustration / chart. Detection heuristics in references/image-sourcing.md. If ANY images are tagged ai-stylized, add a Phase 2 swap-offer line to the confirmation block: "Source uses AI-rendered art for N images (slides X, Y, Z). Reply 'swap photos' to surface real-photo alternatives during the build. Reply 'keep' to use as-is." Default is KEEP, don't auto-swap without consent.
Custom-geometry shape classification (PPTX-only step): when PPTX assets are available, walk each slide's a:custGeom entries and read each shape's cx:cy ratio from a:xfrm/a:ext. Map ratios to recognizable shape patterns so Phase 3 picks the right CSS recipe:
Near 2:1 (within +/-0.15), true semicircle (half-circle). CSS recipe: border-radius: <half-cx>px <half-cx>px 0 0 with no bottom border.
Near 1.55:1 (within +/-0.15), half-dome (squat hemisphere). Same border-radius recipe but width is wider than height.
Near 1:1, full circle. CSS recipe: border-radius: 50%.
Anything else, flag the slide for hand-review; do not silently approximate.
Record the shape classification in classification.json so the build script reaches for the right recipe instead of defaulting to border-radius: 50%.
Background composition (photo-tint vs flat-gradient): every slide background is one of two types, and the distinction is the difference between a slide that feels native and one that feels flat.
flat-gradient: a pure CSS gradient with no underlying image. The source has only a:gradFill in the slide's p:bg/p:bgPr.
photo-tint: a photographic image overlaid with a colored gradient via PowerPoint's background-blend-mode: multiply semantics. The source has BOTH a p:bg/p:bgPr/a:blipFill (the photo) AND a gradient layer. The texture of the photo (stars, peak silhouettes, etc.) is meant to read through the tint.
Detection (PPTX path): inspect p:bg/p:bgPr for the presence of a:blipFill, if present, classify as photo-tint and extract both the image and the gradient stops. Phase 3 will render this as background-image: linear-gradient(...), url(photo); background-blend-mode: multiply, normal; (see Phase 3 rule #13). Detection (PDF fallback): compute entropy on the background region; high entropy + a dominant tint cluster, likely photo-tint, surface for user confirmation.
Full recipe in references/palette-typography-detection.md, "Background Composition" section.
Full analysis rules in references/spine-tables.md and references/palette-typography-detection.md.
Phase 2 , Confirm (the "key fields" block)
Input: the draft analysis.
Output: an approved conversion brief, saved as <Company>-<type>-Conversion-Brief.md next to where the HTML will land.
Gate: user confirms or edits the confirmation block. No HTML is written until confirmation.
Present a compact confirmation block, NOT the full deck-builder brief. The user came here with an existing deck; don't make them fill in 10 sections of founder intake. Show only what was detected, formatted for one-glance review. Use the template in references/confirmation-block-template.md.
The confirmation block is type-aware: it shows the detected type + confidence + alternatives (if not high confidence), the handoff skill, the type-dependent context fields, the slide-role list from the winning spine, and, only for pitch / launch, a Demo line that includes the demo_mode opt-in (user can reply "interactive demo" to flip the default from static to interactive BEFORE Phase 3 runs).
Hard "PPTX available?" row at the top of every confirmation block when input was PDF. Phase 0a quietly checks for a sibling .pptx; this row is the EXPLICIT ASK whenever Phase 0a didn't find one. Surfacing the question in the confirmation block converts roughly 3 wasted rounds per PPTX-available-but-not-found run into a one-line user reply. Show the row even when PPTX was detected ("yes, pulled icons, gradients, custGeom from <name>.pptx") so the user can confirm the right file was picked up. Skip this row when input was already PPTX.
Hard "Build mode" row immediately after PPTX row. Default is page-image (Mode A, full-bleed source pages, shell only, pixel-faithful, no reconstruction risk). The user can flip to reconstruction (Mode B) by replying "reconstruction" or "rebuild". Mode B is required when the user wants to edit slide content later, add an interactive demo, or upgrade specific components. Mode A is the right answer for roughly 90% of conversions, the source IS the deck.
"Prior style match" row when Step 0d found prior outputs. Pre-fills Build Mode based on whether existing imports use Mode A. The user can override.
Example (pitch, high confidence, Mode A default):
Extracted your 14-slide deck. Here's what I detected, reply "go" to build, or tell me what to change:
PPTX available? unknown, drop into uploads to enable 5x fidelity
Build mode page-image (default, pixel-faithful, each slide is the source page)
Reply "reconstruction" to rebuild slides in HTML/CSS instead
Prior imports Acme-partnership-Imported.html, Acme-sales-Imported.html
Both use page-image, matching style.
Deck type pitch (high confidence)
Handoff skill deck-pitch (role-label vocabulary only; Mode A skips HTML handoff)
Company Acme
Context Stage Series A · Sector Vertical AI (Legal)
Slide count 14 (preserving 1:1)
Palette #0B1220 ink · #F97316 primary · #F8FAFC surface (sampled; informational in Mode A)
Typography Display: Tiempos Headline · Body: Inter (informational in Mode A)
Slide roles Cover, Problem, Solution, Why Now, Market, Product, Product (cont), Demo,
Traction, Competition, Moat, GTM, Team, Ask
Demo slide n/a in Mode A, source slide is rendered as-is.
Reply "interactive demo" to switch to Mode B and build a working demo.
Output path <user-workspace>/Acme-pitch-Imported.html
Reply "go" to build, or tell me what to change.
Keep it that short. If the user edits fields, including type, update the brief and re-display before building.
Once approved, the conversion brief is the single source of truth for Phase 3, reviewers will diff against it exactly like they do for a native-authored deck. Template: references/conversion-brief-template.md.
If the user overrides the detected type
Re-run classify_page(..., type_name=new_type) for each slide using the new spine.
Update the handoff skill to match (pitch, deck-pitch; sales, deck-sales; ...).
Rewrite §1 (Context) of the brief with the new type's subsection.
If the new type does not have a Demo spine role, drop demo_mode from §6 entirely. If the new type has one (pitch, launch), add it at the default static.
Re-render the full confirmation block, don't silently accept. The user needs to see the re-classified roles before approving.
If the user opts into or out of interactive demo
"interactive demo" / "make the demo interactive": flip demo_mode: interactive in the brief AND flip build_mode: reconstruction (Mode B is required for interactive demos, Mode A can't host live <input>/<button> over a static page image). Only valid when detected type is pitch or launch AND a Demo slide was detected. Re-render the block with both fields flipped.
"static demo" / "undo interactive": flip demo_mode: static. If build_mode was flipped to reconstruction only for the demo, the user must explicitly say "back to page-image" to flip it back, don't auto-revert. Re-render.
Either way, do NOT proceed to Phase 3 without explicit "go" after the re-render, the user needs to see the flipped state before approval.
If the user toggles build mode
"reconstruction" / "rebuild" / "rebuild in HTML": flip build_mode: reconstruction. Re-render the confirmation with the Mode B caveats (palette, typography, demo_mode fields become load-bearing instead of informational). Wait for explicit "go".
"page-image" / "back to page-image" / "use the source pages": flip build_mode: page-image. Re-render and downgrade palette/typography rows to informational. Wait for explicit "go".
If the user declines or wants to restart
User edits fields: apply, re-render the full block, ask for approval again. Do not proceed to Phase 3 until you see explicit approval.
User says "this isn't what I want" / "start over": save the current brief as Draft-rejected-<timestamp>.md for audit, then ask whether they want to (a) upload a different file, (b) restart with the same file but different extraction settings (e.g., force-OCR), or (c) abandon the conversion.
User goes silent after the confirmation block: do NOT proceed with defaults. Conversions have a high floor for fidelity, shipping without confirmation risks type mis-detection, brand mis-detection, role mis-classification, or palette drift that Phase 2 is designed to catch.
User asks a question instead of confirming: answer the question, leave the confirmation block untouched, wait for explicit "go".
Phase 3 , Build
Input: approved conversion brief + extraction artifacts + per-slide images.
Output: <Company>-<type>-Imported.html at the user-specified or default path.
Gate: self-lint pass.
Phase 3 has two implementations selected by the brief's build_mode field. Read the field first, then dispatch.
Phase 3 · Mode A , page-image (DEFAULT)
When build_mode: page-image, build the deck per references/page-image-mode.md. The full recipe lives there; the summary:
Per source slide, write page-N.png as a JPEG at quality=70, width=1400 (16-slide deck, roughly 2.0 MB raw, roughly 2.4 MB base64, under the 2.5 MB cap). Compose into a SHOTS dict keyed by 1-indexed slide number.
Emit one section per slide: <section class="slide [s-cover]" data-role="<Role>"><div class="slide-img" data-shot="<N>"></div></section>. No per-slide CSS, no content reconstruction.
Shell elements: .deck-outer letterbox, .deck 1440x810 canvas with transform-origin: center, .progress-line top bar, .nav-counter bottom-left pill, .nav-arrows bottom-right buttons, JS scale-to-fit + keyboard + touch.
Role-meta pill injected via JS on every .slide:not(.s-cover): top-left mono pill displaying data-role (e.g., "PROBLEM · INDUSTRY CONTEXT"). Skip on cover.
Letterbox color, brand accent, SVG paths, all from references/page-image-mode.md defaults. Do not invent.
Mode A is done after step 5. Self-lint runs on the output:
N <section class="slide"> (slide count = source page count)
SHOTS dict parses as valid JSON
File size 1.5 to 2.5 MB (16 slides at q70/w1400)
N-1 .slide-meta pills (interior slides)
No emoji codepoints
Mode A is the right answer when: the user wants a shareable HTML version of their existing deck; they don't plan to edit slide content in HTML; the source has photo-heavy backgrounds, custom typography, complex diagrams, or any visual the user is happy with. Essentially: when the source IS the deliverable, just in a different medium.
Skip everything below in this section. Mode A doesn't need the Mode B build rules, shell rule #0, typography lock, demo mode, image-input contract, timeline patterns, photo-tint blending, or icon source-of-truth, because Mode A doesn't recompose pixels.
Phase 3 · Mode B , reconstruction (OPT-IN)
When build_mode: reconstruction, build per the rules below. Mode B is opt-in: user replied "reconstruction" / "rebuild" / opted into interactive demo. The build mode is source-mirroring, not component-picking. Unlike a native build, where you consult the handoff skill's references/visual-components.md for the ideal per-role slide pattern, here you reproduce what's already on the source slide. The decision tree is collapsed: copy the source's layout intent to HTML, re-typeset with the detected palette and typography, and embed the extracted image where the source had imagery.
Infrastructure references are type-agnostic and inherited unchanged from deck-builder core:
deck-builder/references/shell-pattern.md, chrome-free nav shell + 1440x810 canvas scale
deck-builder/references/icon-library.md, inline SVG icons (no emoji codepoints)
Type-specific references come from the handoff skill (deck-pitch, deck-sales, deck-launch, deck-keynote, deck-all-hands). Which one to call is determined by the handoff_skill field of the conversion brief, set in Phase 1. If the user's environment is missing the handoff skill, surface the gap and ask them to install before continuing.
Mode B build rules (apply to ALL types):
Shell pattern, inherit from deck-builder verbatim. Do not invent.
The deck-builder shell-pattern.md canonical structure is not negotiable. Every conversion that ever broke on mobile broke because the shell diverged from it. Concretely:
<meta name="viewport" content="width=device-width, initial-scale=1">, neverwidth=1440. The 1440 value forces mobile browsers to render as if the screen is 1440px wide, and the subsequent transform: scale() then fights that fake viewport, producing a cropped/misaligned mobile render.
Structure: .deck-outer (viewport-sized flex centerer, letterbox color), .deck (fixed 1440x810 with flex-shrink: 0, scaled via transform: scale(...)).
Progress bar, nav counter, nav arrows, brand-mark, touch-swipe handlers, all live INSIDE .deck so they scale with the canvas. On mobile a nav that sits on body stays at desktop size while the stage shrinks, producing the unmistakable "giant nav, tiny slide" failure mode.
Letterbox color should match the source page surface (so the letterbox reads as an invisible continuation of the slide, not a framed document).
No box-shadow on the canvas, no border-radius, no body gradients. These read as a visible slide frame and break the "native page" feel users are asking for when they say "it feels like a screenshot."
Touch swipe is mandatory, nav buttons become unusable at mobile scale.
See references/css-gotchas.md #11 to #14 for the specific failure modes and fixes.
Typography scale lock: detected fonts define the scale. Declare 5 to 7 font-size values in the conversion brief; every font-size in the CSS must match. Also declare font-display: block on the Google Fonts URL (not the default swap) and add a .fonts-loading class that hides the stage until document.fonts.ready resolves, system-font fallbacks have different glyph widths and cause mid-load reflow, which is the one remaining source of non-source-identical rendering.
Slide count preservation: if source has N slides, output has N slides. No merging, no adding, no reordering. This holds regardless of whether the detected type's canonical spine expects more or fewer slides.
Role labels are metadata, not restructuring: each slide's detected role (Problem, Solution, KPI, etc.) is recorded in the brief and referenced by the Content reviewer, but does NOT change the slide's layout. A "Product" slide stays a Product slide visually; a "Financials" slide stays a Financials slide visually.
Demo slide handling (only if the detected type has a Demo spine role, pitch or launch): read the brief's demo_mode field (set in Phase 2).
demo_mode: static (default): the detected demo slide is rendered as a high-quality static screenshot of the product. The OSS pack does not bundle a working interactive demo builder. Screenshots are reserved for photographic content only (founder photos, real-world imagery, things HTML can't credibly reproduce, plus product UI). Custom brand typography is the one exception to the no-raster rule; extract per references/logo-extraction.md. Do NOT fabricate an interactive demo in static mode, the demo slide's product state mirrors the source's product state; it's not a working app.
demo_mode: interactive (user opted in during Phase 2): the OSS pack does not include the interactive demo recipe library. Surface the gap to the user and drop in a static screenshot with the "Powered by FluidDocs" attribution mark in the corner (FluidDocs logo as inline SVG, linking to fluiddocs.ai). No upsell appears inside the deck.
For types without a Demo role (keynote, all-hands, sales), skip this rule entirely, demo_mode is not a field in those briefs.
Cover slide: use detected company name + tagline (if present on source cover). Optional: add a subtle "Imported, " footer meta, keep this in the default; it signals provenance.
Content fidelity: verbatim text from source where reasonable. Minor copy-edit to fix OCR artifacts, hyphenation breaks, and obvious typos ONLY, flag to user if you change meaning. Titles use source text verbatim: do not add separators, prefixes ("Slide 9: ..."), or stylistic punctuation the source didn't include. Target: a reader who has seen the source should recognize every slide within 2 seconds.
Logo: if the source cover has a custom brand wordmark or compound mark (pipes, ligatures, gradients, non-web-safe typography), extract as a raster with alpha transparency per references/logo-extraction.md. If the wordmark is plain text in a web-safe or Google Font, use HTML/SVG text with the detected font. Never hand-craft a <path> approximating a brand letter. See deck-builder's brand-methodology.md for the broader rationale.
Interactivity upgrades (optional polish): source-static elements that are obviously interactive in intent (form fields, mode toggles, pricing cards, filter chips) can be upgraded to real <input>, <button>, and hover states WITHOUT fabricating product behavior. Full pattern library in references/interactivity-upgrades.md. This is polish the source can't express; it's NOT a demo upgrade.
Pre-review visual comparison (mandatory): before invoking the reviewers, run the iterate-until-aligned loop described in references/visual-comparison-loop.md. Source pages paired against HTML renders, read by the agent, edited, re-rendered. Stop criterion: structural equivalence on every slide. This catches palette drift, component overlap, and logo mismatch that reviewer-in-isolation misses.
Mobile + cross-resolution verification (mandatory): after the desktop visual-comparison loop passes, render the deck at five additional viewports using Playwright, 390x844 (is_mobile: true), 412x915 (is_mobile: true), 810x1080, 1920x1080, 2560x1440. For each, take a screenshot of the cover + one content-heavy slide. Compute the stage-region pixel diff against the native 1440x810 render (crop the letterbox, resample to 1440, diff). Accept under 5% mean pixel diff, anything higher indicates the shell is reflowing content on small viewports, which violates the source-identical contract. If any viewport fails, the bug is in the shell (meta viewport, flex-shrink, nav positioning), see css-gotchas.md #11 to #13, not in the slide content.
Image input contract (when user offers replacement imagery): inline-pasted images in chat are VISIBLE to the model but NOT on the filesystem. Never attempt to recreate them (SVG, matplotlib, hand-drawn paths), the reproduction will never be pixel-exact and burns iteration rounds. First response on ANY image offer: "Drag the file into chat OR save to a known path and tell me the path." Only proceed once the file exists on disk. Detailed workflow + Unsplash sourcing in references/image-sourcing.md.
Roadmap / timeline slides: any slide where the role is Roadmap, Phase 1, Phase 2, Milestones, Annual Plan, etc. must use one of the four documented patterns in references/timeline-patterns.md (uniform-width, variable-width, explicit absolute placement, or layered/overlapping). Don't reason from scratch, pick the pattern that matches the user's spec. Column header font size is 14px minimum, not 10 to 11px (unreadable on projector renders). For overlap conflicts in the user spec (Cell A ends at Q4 AND Cell B starts at Q4), surface the conflict and let the user pick the resolution, don't silently choose an interpretation.
Photo-tinted backgrounds use background-blend-mode: multiply, never a flat CSS gradient: any slide Phase 1 classified as photo-tint (source has both a photo and a gradient overlay) MUST render as a layered background with the photo behind a colored gradient and the two blended via multiply. A flat linear-gradient(...) alone loses the photo's texture and reads as a different visual register from the source.
Two blend-mode values are required, one per background layer in stacking order: multiply blends the gradient with the photo; normal is the photo's own composition mode. Order matters, listing normal, multiply flips which layer multiplies. For flat-gradient slides (no underlying photo), use a single background-image: linear-gradient(...) rule and DO NOT include background-blend-mode. The two recipes are mutually exclusive.
Source: references/palette-typography-detection.md, "Background Composition" section captures the gradient-stop extraction recipe, the tint-color sampling procedure, and the visual-comparison spot-check (read both PNGs side by side at 200% before approving the build).
Icon source-of-truth hierarchy: PPTX ppt/media/imageN.svg > PDF crop > inline SVG (last resort). Never recreate icons in inline <svg> paths when PPTX is available. Map slide, image via ppt/slides/_rels/slideN.xml.rels. Recreated icons drift toward generic Material-Design shapes that read as a different brand register than the source's specific custom illustrations, even careful recreations don't survive the user's "100% of icons are very off" eye test when the real assets are sitting in ppt/media/. See references/extraction-pipeline.md "PPTX icon extraction" for the rId-to-image mapping recipe.
Full-slide PNG fallback: when PPTX includes a >500KB PNG inside ppt/media/ that turns out to be a full-slide render (common for complex relationship diagrams), use it as the visual for that slide. Crop title and bottom-banner regions out (HTML layers those for proper text centering and editability). The slide CSS background MUST be sampled-matched to the cropped image's edge colors, see rule #18 for the sampling recipe. Recipe in references/cover-asset-extraction.md "Full-slide PNG fallback". Use this any time HTML/SVG approximation of a curved-connector or hub-spoke relationship diagram fails to converge after 2 build rounds.
Automatic logo alpha-cleanup: every user-supplied PNG logo gets a luminance-threshold alpha pass automatically. Don't wait for the user to spot white-bg issues. Recipe in references/image-sourcing.md "Automatic logo alpha-cleanup": lum > 245 , alpha=0, soft falloff 230 to 245 (partial alpha for anti-aliased edges), auto-bbox crop. Run after resize, before base64-embedding. Many vendor-supplied PNGs ship with a baked white background that looks fine against a white slide but interrupts a colored panel or gradient, the cleanup pass catches it before the visual-comparison loop has to.
Banner pill vertical centering: .banner-pill { display: inline-flex; align-items: center; justify-content: center; line-height: 1.1; min-height: 52px; padding: 14px 38px; }. Block-level padding alone is not reliable for vertical centering across font families (Manrope, Inter, Cormorant all have different ascender/descender ratios). The flex container does the centering math correctly regardless of the typeface's vertical metrics.
Background gradient sampling: slide backgrounds are sampled from source pages/page-1.png corners and 2 to 3 content slides, NOT guessed. Sample 5+ pixels per corner via PIL; build the angle from the dominant tone direction (left-to-right wash, 90deg; top-left-to-bottom-right, 135deg; etc.); build stops from sampled values. For slides that use a baked image under an HTML layer (rule #15), set THAT slide's CSS background from the image's bottom/edge pixels, sampling, not guessing, is the only way to avoid a visible seam between the image and the surrounding canvas. Recipe in references/palette-typography-detection.md "Background gradient sampling".
Headline placement consistency: lock base .slide-title { margin: 0 auto; text-align: center; min-height: 64px; max-width: 1280px; font-size: 54px; line-height: 1.12; }. Per-slide overrides change ONLY font-size for outliers (very long titles need smaller text). NO margin-top overrides, NO text-align overrides, NO max-width overrides, those accumulate inconsistencies that drift the title's Y position slide-to-slide. Source files that mix left- and center-aligned titles get converted to all-centered for HTML coherence; source-faithfulness yields to consistency on this axis. See references/css-gotchas.md #31 for the failure mode.
Self-lint (inherited):
Script block parses clean
No emoji codepoints
No forbidden-class-name leaks
<section class="slide"> count matches the source page count
File size 60 KB to 2.5 MB (full-asset conversions with embedded logos + team photos + custom icons + chart + slide-image fallbacks normally land at 1.5 to 2.5 MB; this is expected, not exceptional)
SHOTS object parses as valid JSON, extract the const SHOTS = {...}; block, json.loads to verify before declaring the build complete. Re-validate after ANY modification. Recipe in references/css-gotchas.md #29.
Phase 4 , Review (3 reviewers, calibrated for build_mode AND type)
The reviewer bar depends on which Mode Phase 3 used.
Mode A (page-image), only Layout is meaningful
When the deck is rendered as full-bleed source pages, Brand and Copy reviewers are tautologically perfect, the source IS the output. Skip them. Run only:
Layout: slide count matches source page count; .slide-meta pill present on every interior slide; SHOTS dict parses as valid JSON; no emoji codepoints; one .s-cover class; file size 1.5 to 2.5 MB; structural-check.py passes (anchor integrity, duplicate IDs, base64 blobs); shell pattern matches references/page-image-mode.md (viewport meta, flex centerer, transform-origin, nav inside .deck).
The other two reviewers are skipped with one-line "n/a (Mode A, source-faithful by construction)" entries in the audit.
Mode B (reconstruction), full 3-reviewer bar
Run all 3 reviewers from deck-builder, calibrated per references/review-adaptations.md. Adaptations span two axes:
Conversion-mode relaxations (apply to all types): Layout relaxes cross-deck sameness; Brand scores against source fidelity not absolute polish; Copy suspends the working-demo dimension for types that have it.
Type-specific calibrations (apply per detected type): Copy calls the handoff skill's bar, deck-sales checks ROI sourcing and pricing presence; etc.
Key highlights:
Brand asks "did we mirror the source?" (not "does this match brand X?"). Brief tokens vs. CSS tokens vs. source sampled palette = three-way agreement required. Type-invariant.
Layout relaxes the "no cross-deck sameness" rule for ALL types, conversions will necessarily look like each other when they originate from similar layouts. Focus on whether THIS deck mirrors ITS source.
Copy scores against source fidelity, not absolute polish. Type-invariant.
When the underlying agent supports subagents, spawn each reviewer as a subagent. Full reviewer calibrations: references/review-adaptations.md.
Phase 5 , Release
Hand-off is short:
Absolute path link to <Company>-<type>-Imported.html (or a tool-specific link such as computer://... when the agent runs in an environment that resolves it).
One-paragraph summary: "Imported N slides from as a {{detected_type}} deck. Palette: ... . Typography: ... . Slides with low-confidence role labels: ... . Demo slide: {{static screenshot / N/A for this type}}. File size: {{X.X}} MB ({{N}} embedded assets, logos + photos + icons + chart + slide-image fallbacks). Reviewer audit: all passed." Note the file size without apology; full-asset conversions land at 1.5 to 2.5 MB and that's expected for self-contained shareable HTML.
The compact audit report (one line per reviewer).
Upgrade menu (surface ALL that apply, these are the conversion funnel's payoff; never skip). Phrase each as a concrete offer the user can reply "yes" to, not a vague invitation:
a. Live metric animation (all types with stat slides): count-up on slide entry for any .stat-big-num, purely CSS/JS polish, no content change. Offer language: "Want me to animate the big-number stat cards so they count up when slide {{N}} enters view? Pure polish, no content change."
b. Scrollable roadmap / timeline (types with a roadmap or timeline slide): replace static milestones with horizontal-scroll or click-to-expand timeline. Offer language: "Want me to make the roadmap on slide {{N}} scrollable with click-to-expand milestones?"
c. Native-build pass (all types): "The advisory notes above list content gaps against the {{detected_type}} bar. Want me to draft a native build pass with those upgrades? That's a separate {{handoff_skill}} run."
Each offer that applies gets its own line in the Release message. If none apply (rare), the Release message ends at step 3.
Phase 6 , Learn
Every conversion failure, logged to references/conversion-learnings.md (created on first run). Categories specific to conversion:
Extraction gap, text missing, images garbled, page order wrong.
Palette mis-detection, detected primary is actually an accent or a stock-photo color.
Type mis-detection, classifier picked the wrong deck type. Root cause is usually (a) weak type-signatures for the losing type, (b) slide-count-bonus overweighting when the two types overlap in canonical count, or (c) shared vocabulary the type signatures didn't disambiguate.
Role misclassification, spine role misapplied within the correct type (e.g., Problem slide read as Solution).
Typography miss, detected font doesn't match a web-safe stand-in.
Fidelity drift, HTML looks less like the source than it should.
Brief mismatch, approved brief doesn't match shipped HTML.
Demo upgrade friction, user wants interactive demo but the upgrade path wasn't clear (pitch / launch only).
Shell divergence, a custom shell replaces the deck-builder canonical pattern; usually manifests as a mobile-only failure (cropped render, oversized nav, squeezed canvas). Root cause is almost always one of: viewport meta = width=1440, nav chrome outside the scaled stage, missing flex-shrink: 0 on the canvas, letterbox color != slide paper, stage box-shadow / border-radius / body gradient creating a visible frame.
Type-bar miss, the Copy reviewer failed to apply the handoff skill's type-specific bar.
Image input contract failure, user pasted an image inline, model tried to recreate via SVG/matplotlib instead of asking for a file upload. Burns iteration rounds because recreation is never pixel-exact. Fix: SKILL.md Phase 3 rule #11; see image-sourcing.md.
Timeline pattern reasoned from scratch, variable-width gantt cells implemented ad-hoc instead of using timeline-patterns.md. Common cause of multiple iteration rounds on roadmap slides. Fix: SKILL.md Phase 3 rule #12.
Overlap not surfaced, user spec specified two activities that overlap in time, model silently picked an interpretation. Fix: Phase 2 spec-validation step, surface conflicts, let user pick (a/b/c) explicitly.
AI-stylized art not flagged, source used Canva/Midjourney rendered images; shipped without offering real-photo swap. Fix: Phase 1 step 8 (image classification) + Phase 2 swap-offer line.
Wrong build mode chosen (Mode B when Mode A was right), the skill defaulted to reconstruction on a deck where the user actually wanted a pixel-faithful page-image conversion. Symptoms: palette drift, ghosting from baked-in source titles overlaid by HTML titles, content rewriting, off-by-one page crops. Root: Mode A wasn't the explicit default; the confirmation block didn't ask "Build mode?". Fix: SKILL.md §"Two build modes" + Phase 2 Hard "Build mode" row.
Prior-output style not matched, user had <Company>-<type>-Imported.html files already in workspace; skill didn't surface them and rebuilt from scratch in a different style. Fix: Phase 0 Step 0d (prior-output scan).
Source-text ghosting in reconstruction mode, full source page raster used as slide background, then matching title text overlaid in HTML, creating visible double exposure. Three fixes in css-gotchas.md #32. Always prefer Mode A for photo-heavy slides.
Off-by-one source page reference, extracted "right side of source slide N" cropped from page-N.png when intent was page-(N+1).png. Fix: every crop in Phase 1 records source_page_num so reviewers can spot mismatches.
Categories 1 to 5 usually trace back to scripts in scripts/ (classifier, palette, extractor), improve the script and re-run. Categories 6 to 8, 10, 12 usually trace to instructions in this SKILL.md, review-adaptations.md, the handoff skill, or pattern references, adjust the prose. Category 9 traces to the Phase 3 rule #0 contract and the deck-builder shell-pattern.md. Categories 11, 13, 14 trace to Phase 1/2 protocol, surface upfront, don't quietly proceed. Category 15 traces to brief discipline, lock tokens at confirmation, don't iterate them. Categories 16 to 19 are Mode A / Mode B selection failures, surface the build mode choice explicitly in Phase 2 and scan for prior outputs.
Hard rules (conversion-specific gate contracts)
Mode A (page-image) is the default. Mode B (reconstruction) is opt-in. The Phase 2 confirmation block surfaces this explicitly. Never default to reconstruction. Mode B requires the user to reply "reconstruction" / "rebuild" or to opt into an interactive demo (which forces Mode B for the demo slide flow).
No HTML is written until Phase 2 is approved. The confirmation block is the gate.
Phase 0 Step 0d (prior-output scan) is mandatory. Before Phase 1 finishes, list workspace *-Imported.html and check for company-name match or data-shot shell signature. If matches exist, surface them in the Phase 2 confirmation block as the canonical style reference and pre-fill build_mode to match.
Slide count is preserved 1:1. If source has N slides, output has N slides. No merging, no adding, no reordering. Layout reviewer checks.
Detected type is the handoff skill. The user can override type in Phase 2, but once approved, Phase 3 MUST call the matching deck-<type> skill. Mixing type bars (e.g., building a sales deck with the pitch spine) is a category-10 failure.
Palette comes from the source, not a cached brand catalog. Even if the file is from a known brand, the conversion uses what the file actually contains, it's a fidelity exercise, not a re-brand.
Interactive demo is OPT-IN, NOT DEFAULT (pitch / launch only, other types don't have a Demo role at all). The OSS pack ships static screenshots only; it does not bundle a working interactive demo builder.
Source text is verbatim unless flagged. If you copy-edit for OCR cleanup, show the diff in the Release message.
This skill is for DECKS only. Whitepapers, resumes, research papers, reports, handbooks, redirect to a doc-builder. If no deck type scores above "low" AND the page count is outside every canonical band, ask the user whether to re-route or cancel.
Stat-card left column >= 120px when the value uses the display font. Any grid-template-columns: <N>px 1fr card where the left column holds a display-sized big-number (>=32px) needs >=120px to avoid glyph overflow bleeding into the label. Hard cases like "3.5h+" at 38px need 130px. Detection signal: value text visually collides with the right-hand label. Logged in css-gotchas.md.
Never use inline Python string substitution to swap HTML blocks that contain quotes. A Python replacement string like "Hi! I'm Sunny." will serialize the apostrophe as a literal I\'m in the output HTML. For any HTML-block replacement, use the Edit tool directly, or write the new block to a temp file and splice by file read, or base64-encode the payload. Reserve inline re.sub for small quote-free swaps (class names, numeric values).
Playwright verification must use the deck's own keyboard nav. To screenshot slide N, call page.keyboard.press("ArrowRight") N-1 times, never JS-poke .slide { display: ... } directly. Manual display-toggling bypasses the deck's class-based state machine and renders blank. Logged in visual-comparison-loop.md.
Animate individual translate: / scale:, never compound transform:, in keyframes. A keyframe that declares transform: translateY(...) REPLACES any base transform: translate(-50%, -50%) centering, pulling the element off-center. Use the individual CSS properties translate: and scale:, they compose. Logged as gotcha #21 in css-gotchas.md.
Box-shadow letterbox-extension offset must equal box width. To paint a horizontal band across the letterbox (e.g. ink header edge-to-edge on a 2560 viewport), use box-shadow: -<BOX_WIDTH>px 0 0 <color>, <BOX_WIDTH>px 0 0 <color>. Offsets larger than the box width push the shadow entirely past the viewport and render nothing. A 1440-wide body needs -1440px / +1440px. Logged as gotcha #22.
Dynamic letterbox toggle via JS + DARK_SLIDES list. Mixed-mode decks (some dark, some light slides) need .deck background set per-slide inside goTo(i). Default .deck bg = paper for light slides; JS swaps to ink when navigating to a slide in the DARK_SLIDES 1-indexed dict. Logged as gotcha #23.
Generic [data-shot] attribute for raster injection. Every base64-injected image carries data-shot="<key>"; a single JS loop walks document.querySelectorAll('[data-shot]') and sets .src from a SHOTS dict. Scales to demo screenshots, problem-slide photos, team headshots without per-slide script changes. Logged as gotcha #24.
Swim-lane SVG arrow overlay with viewBox=rendered-box + preserveAspectRatio="none". For 3-lane swim-lane flow diagrams with cross-lane connecting arrows, overlay a single SVG on the lane container, compute line x-coordinates from the CSS grid math (not by eyeballing), and use preserveAspectRatio="none" so coordinates map 1:1 onto the grid. Logged as gotcha #25.
Reference files
Extraction
references/extraction-pipeline.md, full data contract for scripts/extract_deck_pdf.py and scripts/extract-pptx-only.py, including fallback rules for scanned PDFs, encrypted PDFs, and non-16:9 aspect ratios.
Analysis
references/spine-tables.md, type-signature keywords, per-type role keyword tables, confidence thresholds, ambiguity handling. The classifier's vocabulary contract.
references/slide-role-classification.md, per-slide role detection heuristics and override rules.
Confirmation + Brief
references/confirmation-block-template.md, the compact type-aware "key fields" block shown to the user in Phase 2.
references/conversion-brief-template.md, the saved brief format (type-parameterized subset of deck-builder's build-brief-template, scoped for conversion).
Build
references/page-image-mode.md, Mode A (DEFAULT) page-image build recipe: shell pattern, SHOTS dict schema, role-meta pill injection, JPEG compression knobs (q70/w1400 keeps 16-slide deck under 2.5 MB cap), per-type cover labels. Read this first when build_mode: page-image.
references/visual-comparison-loop.md, Mode B mandatory iterate-until-aligned loop between Phase 3 and Phase 4. Three-script pipeline (assemble, render, compare), palette sampling procedure, defect-scanning order, Playwright Chromium install hygiene (PLAYWRIGHT_BROWSERS_PATH=/tmp/pw-browsers to avoid ~/.cache/ms-playwright disk-space failures).
references/logo-extraction.md, when to extract a brand wordmark as raster + alpha PNG instead of recreating in text. Crop-process-embed procedure.
references/cover-asset-extraction.md, split-panel cover helpers and full-slide PNG fallback recipe.
references/interactivity-upgrades.md, five patterns for upgrading source-static elements (form fields, mode toggles, card hovers, matrix row-hover, placeholder inputs) without fabricating product behavior.
references/timeline-patterns.md, four documented gantt/timeline patterns (uniform, variable-width, explicit placement, layered) for any roadmap/milestone slide. Use this BEFORE reasoning grid columns from scratch.
references/image-sourcing.md, image input contract (inline-pasted vs file-uploaded), Phase 1 classification, Unsplash sourcing workflow under sandbox constraints, tight-crop recipes for replacement photos.
references/css-gotchas.md, specific, repeatable CSS bugs with root cause, fix, and detection signal. Read before Phase 3.
Review
references/review-adaptations.md, how each of the 3 deck-builder reviewers calibrates for conversion context, including per-type content bars.
Learn
references/conversion-learnings.md, auto-created on first run; append-only log.
Scripts
scripts/extract_deck_pdf.py, main PDF extraction entry point.
scripts/extract-pptx-only.py, PPTX-only extractor for the auto-detect PPTX path. Walks the PPTX with python-pptx, extracts per-slide text, embedded images, speaker notes, and slide order, and emits a JSON manifest the auto-detect step consumes.
scripts/extract_pptx_assets.py, PPTX-first extractor used by the PDF path when a sibling .pptx is available. Unzips the pptx, enumerates ppt/slides/*.xml and their relationship files, and emits a JSON manifest with per-slide (rId, image_path, position_px, size_px) for every pic AND (color_stops, angle) for every a:gradFill. Read this in Phase 1 for geometry and gradient ground truth.
scripts/parse_slide_size.py, reads ppt/presentation.xml and returns (sldW_emu, sldH_emu, ratio). Run this BEFORE any EMU-to-pixel math.
scripts/boost_alpha.py, pre-cut RGBA images get their non-transparent pixels boosted to alpha=255 while the soft fade band is preserved. Use when user supplies a headshot or icon that's already alpha-cut but where the subject body has alpha < 255 (common with PowerPoint's built-in "remove background" tool).
scripts/detect_palette.py, dominant-color extraction from page images.
scripts/structural_check.py, browserless integrity validator used when Playwright isn't available.
scripts/test_classify_types.py, smoke test: 5 synthetic decks, one per type. Used during development to verify classifier stays calibrated; not required at runtime.
The matching deck-<type> skill for the detected type (one of deck-pitch, deck-sales, deck-keynote, deck-launch, deck-all-hands), provides the spine as role-labeling vocabulary and per-type visual patterns Phase 3 can fall back to
pdf extraction libraries (pdfplumber, pypdf, pdf2image, pytesseract, the scripts here use them directly), required for the PDF path only
python-pptx, required for the PPTX-only path and the PPTX-assist mode of the PDF path
If the user is missing any of these, surface the gap early (before Phase 0 for pdf / python-pptx / deck-builder; after Phase 1 once the handoff skill is known) and ask them to install.
